1. Overview
TA-EDU is an AI tutoring platform for schools and colleges. This policy describes what data we collect, how we use it, who else receives it, and what you can ask us to do with it.
When a school or college deploys TA-EDU, the institution decides what happens to student records and TA-EDU acts on its instructions. The institution’s own privacy policies and notices govern student rights, and its agreement with us governs ours.
Student accounts are anonymous by design. A student account is created without a name or a real email address, and teachers see each student only as “Student N”, a label that is specific to one course. This reduces what we hold about a student. It does not eliminate it: the record of which label belongs to which account exists in our database, and students can still type identifying details into a chat.
2. Data We Collect
- Account information: for teachers, an email address, a hashed password and a role. For students, a generated handle, a hashed one-time recovery code and a role. No student name or personal email is collected.
- Chat messages: the messages exchanged with the AI tutor, including tool output. Before a message is written to our database, identifying spans we can recognise are replaced with a fixed block of characters, and the original is not kept.
- Learning profiles: topic understanding levels, misconceptions and progress, derived by AI from chat interactions.
- Course materials: documents, syllabi, problem sets and other files uploaded by teachers.
- Assessment data: practice attempts and their scores.
- Photos and uploaded work are not stored. A photo a student attaches in chat is read once by the AI to answer that message and is then discarded; only the text of the conversation is saved. Papers a teacher uploads for AI-assisted grading are processed in memory and are not kept.
- Usage data: AI token consumption per request, feature usage and rate limit counters.
- Embeddings: numeric representations of messages and course materials, used for search.
3. How We Use Your Data
- Producing tutoring responses tailored to your course materials and your level.
- Building knowledge profiles so teachers can see where a class is struggling.
- Searching across course materials and past conversations.
- Producing weekly summaries for teachers.
- Enforcing rate limits and accounting for the cost of AI usage.
We do not use student data for targeted advertising, we do not build profiles of students for any purpose other than teaching them in the course they are enrolled in, and we do not sell or rent student data. We do not use student data to train AI models, and neither does our AI provider: see section 4.
4. Who Else Receives Data
Three companies receive data in order for the service to run. We say below what agreement is in place with each, because they are not the same.
- Together AI: every AI feature runs on Together AI, and it is the only AI provider we send anything to. That covers chat messages and the course context sent with them, and uploaded work submitted for grading. It does not cover embeddings, which are computed inside our database provider, Supabase, and are not sent to any AI company. Zero data retention is switched on for our account, which under Together AI’s terms means prompts and responses are not stored, retained, or used for training or product improvement beyond the time needed to answer the request. We have asked Together AI for a data processing agreement and have not yet received one, so today the protection rests on that retention setting and on their terms of service rather than on a separate signed agreement.
- Supabase: stores everything at rest, in a PostgreSQL database with row level security, and computes the embeddings described in section 2 without the text leaving it. Supabase’s data processing addendum applies automatically to our account under its terms of service, and Supabase confirmed that to us in writing.
- Vercel: hosts the application. Requests pass through it and their bodies are visible in transit. Function and request logs are kept for 1 day on our plan. Vercel’s data processing addendum applies to our account. One thing is worth stating plainly: for the usage metadata the platform generates about requests, such as IP address, browser user agent, path and timing, Vercel acts as an independent controller on its own terms rather than on our instructions. That metadata is outside our control and outside the deletion we can carry out for you.
One other service is contacted on a narrower path:
- Hipolabs university directory: used only by the school field on our public request-access form. Our server calls it, not your browser, and it receives only the text typed into that field. No student or teacher session ever reaches it.
Graphs and charts in a tutor answer are computed and drawn inside your own browser by code that ships with the application. Nothing about a graph is sent to a third party.
5. Data Protection
- All data is encrypted in transit via HTTPS and TLS.
- Row level security policies isolate data: students reach only their own, teachers only their own courses.
- Authentication is required for every data access endpoint, and routes check role as well as identity.
- Identifying text in chat messages is blanked out before storage, and a daily pass looks for anything the first pass missed.
- A teacher can read every conversation a student has in that teacher’s course. What they see is the blanked-out stored text under the student’s course pseudonym, such as Student 7, and never a name or an email address.
- Class statistics covering fewer than five students are suppressed, so a single student cannot be identified from an average.
- Rate limiting: 10 requests per minute per student.
- CSRF protection is enforced on operations that change data.
6. Schools, FERPA and California
Where a school or college is subject to FERPA, TA-EDU is built to operate under the school official exception (34 CFR 99.31(a)(1)). The institution names TA-EDU in its annual notification, TA-EDU performs a function the institution would otherwise perform itself, and the institution keeps direct control over how education records are used.
FERPA is not the only framework, and for many K-12 schools it is not the relevant one. For a California school, the Student Online Personal Information Protection Act (Bus. & Prof. Code §22584) applies whether or not FERPA does. Under it we commit that we do not use covered student information for targeted advertising, we do not build a profile of a student except for school purposes, we do not sell or rent covered student information, and we delete a student’s covered information when the school asks us to.
Students at institutions covered by FERPA retain the right to inspect their education records, request amendments, and file complaints with the U.S. Department of Education.
7. Data Retention
Retention is set by the deploying institution’s records policy. By default data is kept for as long as the student is enrolled in the course. We do not currently delete data automatically at the end of a term, so a school that wants it gone should ask.
For a pilot, we keep a course’s student data for up to 14 days after the pilot ends, so that we can review how the tutor behaved, and then we delete it. The school can ask for an export first, or for deletion sooner, at any time, and that request comes first. We carry out this deletion ourselves; it is not automatic.
8. Export and Deletion
You can delete an individual conversation yourself, at any time, from the chat sidebar. For anything wider than that, ask your teacher or your school, and they ask us. There is no self-service button for a full export or a full deletion, and we would rather say so than imply one exists.
When a request reaches us, we can carry out any of the following:
- Export one student: everything we hold about that account, for one course or across all of them, as a single JSON file.
- Export a whole course: the same, for every student in it.
- Delete one student: for one course or for the whole account. This removes conversations and messages, knowledge and mastery records, practice attempts, graded submissions and their uploaded files, tutor assignments, course pseudonyms and enrolments. The account itself is removed when the request covers everything, or when a course-scoped deletion leaves the student enrolled nowhere.
- Delete every student’s data in a course, at the school’s request, along with the class statistics computed from that activity.
Two honest caveats. Records of what a request cost us to run stay behind as accounting totals, with the link to the person removed, so we can still balance our books but cannot trace a figure to a student. And every deletion we carry out is itself logged, with counts and the course it applied to, never with any of the content that was deleted.